Res-U-Net-Based Sleep Arousal Detection Using Limited Polysomnography Channels and Multi-Step Training Techniques

Sleep arousal, a type of sleep disorder, occurs when one wakes up and then goes back to sleep. It is crucial to assess the frequency and duration of sleep arousals to determine sleep quality. Sleep arousal is mainly divided into two categories. The first category is caused by apnea and hypopnea, while the second category is attributed to factors unrelated to apnea and hypopnea. Identifying the latter type can be challenging in a clinical setting as it is relatively concealed. While current methods provide satisfactory results in identifying arousals caused by apnea, the ongoing challenge in sleep research lies in accurately identifying arousals that are not associated with apnea. In this paper, we proposed an automatic non-apnea sleep arousal detection algorithm based on polysomnography (PSG) data. As sleep-related physiological signals impact natural sleep, reducing the number of recording sensors can enhance sleep quality. Thus, we have developed an advanced deep-learning architecture that uses 7 crucial signals from the 13 PSG channels as input. This architecture combines residual blocks with U-Net’s encoder and decoder layers to improve detection accuracy by considering local and global features simultaneously. We used datasets from the Public Cardiology Challenge 2018 to validate our approach. To tackle the problem of imbalanced data, we utilized the Synthetic Minority Over-Sampling Technique (SMOTE) to increase the representation of the minority class. Additionally, we enhanced our model’s performance by incorporating multi-step training techniques. Our results highlight how the algorithm efficiently detects non-apnea arousal using fewer signals. Our algorithm achieved the area under the receiver operating characteristic (AUROC) of 0. S69 and the area under the precision-recall curve (AUPRC) of 0.433, outperforming other studies that did not incorporate electromyogram (EMG) and electrooculogram (EOG) signals.